Abstract: Accurate regional power prediction of wind power clusters is of great significance to the bidding grid on the supply side. Since multiple wind farms in the same area have similar fluctuations under climate influence, they can be regarded as wind farm clusters with temporal and spatial correlation, and the clusters are reasonably divided accordingly. Therefore, a multi-region composite short-term wind power prediction model based on adaptive optimization of affinity propagation（AP） clustering and back-propagation（BP） weighted neural network is proposed. Firstly, the historical data of wind farm clusters are clustered and divided by particle swarm optimization AP clustering method. Then, according to the obtained optimal clustering results, the training set of sub-region samples of wind farm groups is constructed. Finally, the BP neural network based on correlation coefficient weights is used to predict the power of each subregion. The example results show that the proposed method can improve the accuracy of prediction by 1.35% and 2.62% compared with the traditional superposition method and a single BP neural network before 24 h. The results show that the model has superior prediction performance.